Teams running Kubernetes can usually see where they’re overprovisioned. Requests are higher than they need to be, there’s consistent headroom, and capacity sits underused.
This has been true for a while, but it is showing up more often now as more teams run burstier model-serving workloads on Kubernetes and start feeling the cost of overprovisioning more directly.
But those workloads don’t get touched.
This shows up most with HPA-managed services. The inefficiency is obvious; as the HPA scales, the waste scales…
